The development of information technology makes massive data generating at an unprecedented rate, which brings the information overload problem. To address this problem, the recommender system emerges.Recommender systems try to suggest users the potential enjoyed information by analyzing users' characteristic and their historical behaviors. How to model the users' individual preference properly is crucial for them. The most widely used methods are Collaborative Filtering (CF) approaches. Based on the assumption that similar users have similar behaviors on similar items, CF methods aim at predicting users' interests by mining their behavior history. However, they have the data sparsity problem, that is, the behavioral data is typically very sparse and it is indeed hard for traditional CF methods to make accurate recommendation with such insufficient data. With the prevalence of massive web applications, we can deal with diverse kinds of data.For example, the massive content information for users and items, the social relationship between users and user's multiple types of behaviors on the internet (such as the shopping history, reading history and rating history). Their forms are diverse and their attributes are heterogenous. If we can mine these heterogenous information effectively for different recommendation cases, the data sparsity problem will be relieved to improve the recommendation quality. In this dissertation, we will take the research on recommender systems with heterogenous information. For different recommendation problems, we have presented different solutions which emerge different heterogenous information effectively. The main contributions are summarized as follows. 1. We propose a novel method, named Content Topic Feature weighted One-Class Collaborative Filtering, to deal with the implicit feedback data in recommender systems. It attempts to solve the one-class problems of implicit feedback by exploiting the rich content information. Specifically, we get a content topic feature for each user and item to assist distinguishing the potential negative examples from missing data, and extend the Matrix Factorization model by incorporating the content-similarity based weighting scheme. Experiments on real-world data show that the proposed method outperforms state-of-the-art algorithms, which suggests that our method can incorporate the content information into implicit feedback effectively and assist to overcome the one-class problem in thi...
修改评论